Affinda: AI Document Processing Platform

Page contributed by Brianna Charles.

AI document processing platform specializing in resume parsing, invoice processing, and purchase order matching with RAG-powered instant learning capabilities.

Overview

Headquartered in Melbourne, Australia, Affinda serves customers in more than 80 countries across 56 languages. The company evolved from a document parsing specialist into a broader intelligent document processing platform, with continuous platform development spanning 2021 through 2025.

That trajectory accelerated in early 2026 with two product launches: an agentic AI platform with persistent model memory that applies corrections instantly via RAG rather than retraining, and a no-code Integration Agent connecting extracted data to 2,800+ business systems through natural language instructions. An AWS case study confirmed the underlying stack: Amazon Bedrock with Claude Sonnet models as the LLM backbone, Kubernetes on EKS, and SageMaker for custom model training. Andrew Bird, Head of AI, reported a "90 percent reduction in configuration time for new document extraction use cases" after the generative AI transition, along with 90% cost savings for the product delivery team.

Earlier moves built the foundation. In September 2025, Affinda gained recognition as an AI recruitment tool for anonymous application processing, positioned alongside Greenhouse for bias reduction. By January 2026, the company appeared in ResearchAndMarkets analysis of the purchase order matching AI market alongside IBM, Oracle, and SAP SE in a sector projected to reach $4.85 billion by 2029.

Affinda claims greater than 99% extraction accuracy across document types, though independent testing tells a more nuanced story. A March 2026 Koncile benchmark testing invoice OCR across 10 tools found errors in 5 of 30 invoices on key fields (Customer ID/SIRET, total amount), and only 7 of 15 complex invoices produced usable line-item results. Koncile is a competing product, which limits the benchmark's neutrality, but it represents the only published third-party accuracy test available. Three named customer deployments provide operational evidence: StateCover Mutual processing 300,000+ case management documents annually across 80 document types, Northline achieving 82% straight-through processing from day one across 120,000+ logistics documents at 13 depots, and Lockerbie Estate reducing invoice processing time by over 50% while automating three document types without developer resources.

Pricing is usage-based with a free trial covering all features. Deployment spans self-serve, API, and custom solution design, with data centers in Sydney, Frankfurt, and Oregon.

How Affinda processes documents

Affinda's pipeline combines frontier LLMs (Claude Sonnet, GPT-4o, Gemini) hosted on AWS and Azure with fine-tuned smaller models for pre- and post-processing. The architecture layers four components: a proprietary reading order algorithm, Microsoft Azure's OCR engine with layout reconstruction, grounded LLM extraction, and RAG-based model memory using Elasticsearch for vector searches. The RAG layer enables what Affinda calls "persistent model memory," where human corrections are applied instantly without retraining cycles.

Andrew Bird described the difference: "If a human corrects a model in the Affinda Platform for a particular document, this correction is added to the platform's model memory and it instantly learns not to make that mistake again. Systems that rely on fine-tuning smaller models will likely require some number of examples (usually 20+) before the model is likely to learn from its mistakes in the next training run." That competitive claim against fine-tuning-based approaches is testable but has not been independently validated.

The platform's core processing capabilities span RAG model memory for instant learning without retraining, grounded LLM extraction tied to source documents, proprietary OCR with layout reconstruction, Pydantic model validation and TypeScript schema generation, anonymous candidate screening for bias reduction, purchase order matching and 3-way reconciliation, agentic workflow automation for multi-step tasks, and a no-code Integration Agent that generates integration code from natural language to connect 2,800+ enterprise systems.

Affinda's claim of being the "first IDP platform to apply agentic AI" carries no third-party citation. In a market where ABBYY, Hyperscience, and Instabase have each made agentic or LLM-native claims, that framing requires external corroboration before it carries weight. Open-source alternatives such as Unstract also offer LLM-based document processing with no-code configuration, providing a useful comparison point for teams evaluating deployment flexibility.

With Amazon Bedrock, there's very little need for specific model training data. As a result, we've seen a 90 percent reduction in configuration time for new document extraction use cases.

Andrew Bird, Head of AI, Affinda (AWS case study)

Use cases

HR technology and recruitment

Affinda's longest-established use case. The platform handles resume and CV parsing with anonymous screening to reduce hiring bias, a capability that drew recognition alongside Greenhouse in UK recruitment AI analysis in September 2025. Additional capabilities include job-matching data extraction, candidate summaries, and identity document processing.

Financial document processing

Invoice processing, validation, and purchase order matching with 3-way reconciliation form the core of Affinda's financial document offering. The Koncile benchmark provides the most detailed independent look at invoice accuracy: standard invoices performed reasonably (25 of 30 correct on key fields), but complex invoices with multi-line descriptions generated what Koncile called "parasitic lines" making data difficult to standardize. Custom field selection via LLM is available, though line-item customization is not.

The company's appearance in the ResearchAndMarkets purchase order matching analysis alongside IBM, Oracle, and SAP signals growing enterprise recognition in procurement automation. The platform also covers lending and banking KYC and compliance documents. Teams evaluating financial document processing alternatives may also want to review Alkymi, which specializes in extracting and transforming unstructured financial documents into standardized datasets for institutional workflows.

Insurance and claims

StateCover Mutual's deployment (300,000+ case management documents annually across 80 document types) is the platform's most cited insurance reference. StateCover Business Analyst Nick Tran describes "superior accuracy and efficiency" but provides no pre/post performance comparison. The platform also handles ACORD forms and general insurance claims workflows.

Logistics and supply chain

Northline's deployment remains Affinda's strongest operational proof point: 82% straight-through processing from the start of deployment across 120,000+ Proof-of-Delivery, Pallet, and Manual Charge documents annually across 13 depots. Head of Business Systems Jorg Both told idm.net.au: "Achieving 82% straight-through processing from the start has given us a strong platform to extend automation into invoices and dangerous good notifications." The platform also handles bills of lading and broader logistics documentation.

Teams evaluating logistics-specific IDP may also want to review LangExtract, Google's open-source Python library for structured extraction from unstructured text, which takes a different architectural approach to grounding extraction in source documents.

Enterprise automation

The March 2026 Integration Agent launch targets the last-mile problem where extracted data must reach systems of record. General Manager Charlie Bellingham explained the rationale: "We saw an opportunity to develop a flexible and customisable tool that lowers the investment required to get an end-to-end document processing solution off the ground and help more complex organisations prove concepts faster."

Property developer Lockerbie Estate deployed the agent to automate supplier invoice processing into Xero. Finance Manager Louis Wright reported: "We were spending more than half our working week getting supplier invoices and sales quotes into Xero. With Affinda Platform's AI Integration Agent, we had a working integration in days without waiting on a costly, time-consuming IT roadmap."

Multi-format document classification and routing spans PDF, JPG/JPEG, PNG, TIFF, DOCX, Excel/CSV, and HTML inputs across document types not covered by vertical-specific modules. Teams prioritizing no-training deployment may also want to review Workist, a Berlin-based platform that similarly claims no-training implementation for mid-market ERP automation.

Technical specifications

Resources

• Website
• Affinda Platform
• AWS case study
• Quick start guide
• Enterprise security
• Github
• Free trial
• Competitive analysis

Company information

180 Flinders Street
Melbourne VIC 3000 Australia
affinda.com
contact@affinda.com
LinkedIn